When a satellite image classifier trained on one part of the world is deployed on another, accuracy commonly falls by 15 to 30 percent (Position: Mission Critical – Satellite Data is a Distinct Modality in Machine Learning). Self-training is one of the few remedies that helps on real geographic benchmarks (Extending the WILDS Benchmark for Unsupervised Adaptation). In self-training, the model labels unlabeled images from the new region and then retrains on its own guesses. The weak point is deciding which guesses to trust. Current methods keep a guess only when the model's predicted probability is high. That rule is known to be unreliable under geographic shift, because confidence can collapse or inflate across a whole region at once (Trusting Small Training Dataset for Supervised Change Detection).
We propose to choose trusted pseudo-labels, and to seed the new region's class prototypes, using the shape of the unlabeled data's feature distribution rather than the classifier's confidence. The idea is to find "modes": places in feature space where many images bunch together. The twist is to count density only across distinct geographic locations. A bunch of nearly identical tiles from one scene does not make a mode, but a pattern that recurs in many separate places does, even if it is rare overall. Images that belong to no such recurring mode are treated as noise. Images attached to a mode share that mode's label.
Core Claim
Density-based filtering has an ambiguity that the literature has not resolved. Low density marks both corrupted samples and legitimate rare classes. Two lines of work pull in opposite directions. Density-peak seeding places prototypes only at high-density, well-separated points (Clustering by fast search and merge of local density peaks for gene expression microarray data; GDD-K-Means Text Clustering Algorithm Based on Grid Filtering Distance and Density of Outliers). That suppresses outliers but ignores rare classes. Density-Aware Clustering Head Initialization (DACHI) in Samples Are Not Equal: A Sample Selection Approach for Deep Clustering does the reverse: it weights samples inversely to their density so that rare modes get a say. In satellite imagery this would also promote isolated artifacts such as haze, sensor glitches and odd crops.
Our claim is that geographic data supplies a signal that separates the two cases. Under Tobler's law, nearby tiles look alike, so raw feature density in Earth observation is inflated by spatial autocorrelation. Many tiles from one scene crowd together whether they show a real category or one scene's artifact. A genuine land-use category, even a rare one, tends to recur across many unrelated locations. A scene-specific corruption does not. We therefore define density by counting distinct spatial blocks among a sample's neighbors, and we treat a mode as real when it is a local maximum of this geo-diverse density, however low its absolute value. Two predictions follow. Seeding prototypes at such modes, with equal weight per mode, should keep rare classes that density peaks drop. It should also reject artifacts that inverse weighting adopts. Filtering pseudo-labels by attachment to these modes should beat confidence thresholds on target-region accuracy, most of all on rare classes.
The obvious limitation must be stated up front. Some corruptions are not scene-specific. Clouds and haze recur across locations, so geo-diverse density will treat them as real modes. We handle this by labeling modes with the classifier rather than discarding them. We will measure how often artifact modes survive.
Research Program
Take a source-trained encoder $f$ and unlabeled images $x_i$ from a target region, each with a location $g_i$. Partition the region into spatial blocks $B(g)$ of fixed size; the block size is a studied hyperparameter. Let $N_k(i)$ be the $k$ nearest neighbors of $f(x_i)$ among target features, found with an approximate nearest-neighbor index to avoid quadratic cost. The geo-diverse density is
$$\rho^{geo}_i = \big|\{\, B(g_j) : j \in N_k(i),\ B(g_j) \neq B(g_i) \,\}\big|,$$
the number of distinct other blocks represented among the neighbors. Neighbors from the sample's own block are excluded, so a scene cannot vouch for itself.
We then reuse the density-peak construction. For each sample we compute $\delta_i$, the feature distance to the nearest sample with higher $\rho^{geo}$. Modes are samples that are local maxima of $\rho^{geo}$ within their neighborhood and have $\delta_i$ above a threshold. Because modes are defined as local maxima rather than by an absolute density cutoff, a rare class can form a mode at low density. This addresses the varying-density failure noted for global cutoffs.
Each remaining sample $i$ is attached to the mode $m(i)$ reached by following links to its higher-density nearest neighbor. It is kept only if density along that chain never falls below a fraction $\alpha$ of the mode's own density:
$$\min_{j \in \text{path}(i \to m(i))} \rho^{geo}_j \ \ge\ \alpha\, \rho^{geo}_{m(i)}.$$
This threshold is relative to each mode, so sparse rare-class basins keep their members. Samples that fail the test are treated as noise for this round.
Each mode gets a label by majority vote of the source classifier over its core members, meaning the samples above the $\alpha$ level. Two consequences follow. First, a sample's pseudo-label is accepted only when the classifier's prediction agrees with its mode's label. In a second variant, the sample adopts the mode label outright. Second, when the deployment head is re-initialized for the target region, prototypes are averages of mode cores with equal weight per mode. This is neither density-weighted nor inversely density-weighted.
Because features drift during adaptation, modes and attachments are recomputed at the start of each self-training round. A lightweight version recomputes them every epoch on a subsample. Training otherwise follows the Noisy Student recipe benchmarked in WILDS 2.0, so the only change is which pseudo-labels survive and how the head is seeded.
There is also a secondary, label-free question. WILDS 2.0 points out that adaptation methods are usually tuned with target labels that a real deployment would not have. We will test whether the DISCO score from Internal Evaluation of Density-Based Clusterings with Noise, which scores noise assignments by their sparsity and isolation, can choose $k$, $\alpha$ and block size without target labels. DISCO was built for static clusterings. We use it only as a model-selection criterion and report how well it correlates with true target accuracy.
Baselines
The naive reference point is source-only empirical risk minimization (ERM), meaning ordinary training on labeled source regions, with the same backbone and augmentations. WILDS shows that ERM is hard to beat on real shifts (WILDS: A Benchmark of in-the-Wild Distribution Shifts). Any gain must first clear it.
The strongest existing competitor on FMoW-WILDS is Noisy Student self-training with confidence thresholding. WILDS 2.0 identifies it as one of the few methods with clear gains on FMoW. The file gives no FMoW scalar for it: no reported result found on Lacuna for Noisy Student on FMoW-WILDS. The only reported Noisy Student figure is on Camelyon17, where out-of-distribution accuracy rose from 70.8% to 86.7%. We will therefore rerun Noisy Student ourselves using the WILDS 2.0 protocol, unlabeled splits and backbone. It will get the same number of rounds, compute, and hyperparameter grid size as our method. FixMatch, also benchmarked there, joins as a second confidence-based baseline.
On FLAIR, the strongest reported system is GeoMultiTaskNet (GeoMultiTaskNet: remote sensing unsupervised domain adaptation using geographical coordinates). It adds a coordinate-prediction auxiliary task and frequency-based class sampling, and reports 47.22% mIoU on its three held-out departments. We will reproduce it on the same split. Beating it means exceeding its reproduced mIoU, with 47.22% as the published reference. Our method can also be stacked on top of it, and that combination will be reported too.
Three density baselines isolate the key ingredient. The first is DBSCAN-style pseudo-label pruning as in Unsupervised Person Re-Identification with Multi-Label Learning Guided Self-Paced Clustering. The second is per-class density ranking in the spirit of CurriculumNet, computed from predicted rather than true labels. The third is DACHI inverse-density head initialization. The decisive ablation keeps our entire pipeline but replaces $\rho^{geo}$ with ordinary $k$-NN density. If that ablation matches the full method, the geographic ingredient is not doing the work.
A class-balanced confidence baseline is also included. It keeps the top fraction of pseudo-labels per predicted class, which directly targets the rare-class problem without any density estimation. Every method gets the same tuning budget. Following the WILDS 2.0 critique, results are reported twice: once tuned on held-out source regions only, and once with oracle target validation.
Why not just…?
Why not just keep confidence thresholding, which already works on FMoW? It does help, and it is the bar we must clear. The objection to it is concrete. Under regional shift, confidence moves for whole scenes at once: the change-detection study saw its reliability score fall to 0 on Dubai while accuracy collapsed. Density in feature space is a property of the target data itself rather than of the classifier's calibration. We will settle the question with head-to-head target accuracy. We will also measure the precision of accepted pseudo-labels at matched acceptance rates, so the gain cannot come simply from keeping more or fewer samples.
Why not just use class-balanced thresholds to protect rare classes? This is the most serious threat. It is cheap and attacks the same rare-class failure. The difference shows up when the classifier's predicted classes are themselves wrong in the new region. A per-class quota then fills rare-class slots with confidently mislabeled samples, while mode attachment uses agreement with recurring structure. If class-balanced thresholding matches us on rare-class recall and pseudo-label precision, we will report that density adds nothing beyond balancing. We will then retreat to the narrower claim about head initialization.
Why not just use DBSCAN, or DACHI's inverse density, which already exist? Both ignore spatial autocorrelation. DBSCAN uses a global density level, so sparse rare-class clusters become noise. DACHI favors low-density points, including scene-specific artifacts. The ordinary-density ablation and the DACHI baseline measure exactly this. We will also inspect accepted modes by hand on a labeled sample to count how many are artifacts under each scheme.
Why not just condition on coordinates, as GeoMultiTaskNet does? Coordinates tell the model where an image is, not which target images are trustworthy. Escofet et al. found explicit metadata gave negligible gains for large geospatial foundation models (Fine-tune Smarter, Not Harder: Parameter-Efficient Fine-Tuning for Geospatial Foundation Models). Our method uses location only to decorrelate density estimates, so the two approaches are complementary. The stacked FLAIR run tests whether they add up.
Proposed Experiments
The primary testbed is FMoW-WILDS, built on the fMoW release (Functional Map of the World), whose image metadata supplies locations. We use the WILDS 2.0 unlabeled target-region data and report average and worst-region accuracy. The second testbed is FLAIR segmentation with the 10-source and 3-target department split, reporting mIoU. Here density is computed on patch-level features and pseudo-labels are filtered per patch.
A third run on Sen1Floods11's geographic hold-out set uses a LoRA-adapted Prithvi backbone, where LoRA reached 87.57% mIoU against 82.07% for full fine-tuning. It tests whether mode-based filtering still helps when a frozen foundation backbone already limits forgetting. If our method helps on FMoW but not here, that is informative rather than fatal.
Beyond headline accuracy, four diagnostics decide whether the mechanism is real. The first is pseudo-label precision as a function of acceptance rate, for every filter. The second is per-class recall split into frequency terciles, which is where the predicted advantage over density peaks and confidence filtering should concentrate. The third is the fraction of discovered modes that a human audit of 500 sampled tiles labels as artifacts, compared between geo-diverse and ordinary density. The fourth is the rank correlation between DISCO and true target accuracy across the hyperparameter grid.
A controlled stress test injects synthetic scene-specific corruptions into a known fraction of FMoW target scenes: blur, banding and saturated patches applied to whole scenes. It then checks whether $\rho^{geo}$ rejects them while ordinary density keeps them. Running the same test with corruptions spread across many scenes shows where the method stops working.
The proposal succeeds if three things hold together. It must beat our reproduced Noisy Student on FMoW worst-region accuracy under source-only tuning. It must beat the ordinary-density ablation and class-balanced thresholding on low-frequency-class recall. And it must keep fewer artifact modes than DACHI-style weighting. It fails if the ordinary-density ablation performs as well, because then the geographic insight is not the cause of any gain.
References
- Position: Mission Critical – Satellite Data is a Distinct Modality in Machine Learning
- Extending the WILDS Benchmark for Unsupervised Adaptation
- Trusting small training dataset for supervised change detection
- Clustering by fast search and merge of local density peaks for gene expression microarray data
- GDD-K-Means Text Clustering Algorithm Based on Grid Filtering Distance and Density of Outliers
- Samples Are Not Equal: A Sample Selection Approach for Deep Clustering
- Internal Evaluation of Density-Based Clusterings with Noise
- WILDS: A Benchmark of in-the-Wild Distribution Shifts
- GeoMultiTaskNet: remote sensing unsupervised domain adaptation using geographical coordinates
- Unsupervised person re-identification with multi-label learning guided self-paced clustering
- CurriculumNet: Weakly Supervised Learning from Large-Scale Web Images
- Fine-Tune Smarter, Not Harder: Parameter-Efficient Fine-Tuning for Geospatial Foundation Models
- Functional Map of the World